USA_OD_2025 1/9 Arc-Second U.S. Offshore DEMUSA_OD_2025 美国沿海 1/9 弧秒离岸 DEM

Dataset overview

USA_OD_2025 is a 1/9 arc-second offshore digital elevation model for major U.S. coastal zones. The product is generated with DEM-AMSSRN, an asymmetric multi-scale super-resolution network designed to preserve continental-shelf geomorphology while recovering fine-scale coastal terrain detail.

The source data are the CUDEM 1/3 arc-second CONUS: NAVD88 offshore DEM, which provides approximately 10 m input resolution. DEM-AMSSRN enhances this input by a factor of 3 to produce a 1/9 arc-second grid at approximately 3 m resolution. The released product covers the East Coast, West Coast, Gulf Coast and Alaska where source coverage is available.

• CUDEMs provide 1/9 arc-second offshore topobathymetry at approximately 3 m and 1/3 arc-second bathymetry at approximately 10 m
• The source CUDEM collection contains 872 high-resolution offshore tiles, each 8112 x 8112 pixels
• Sixty representative tiles were selected across East, West and Gulf coastal morphologies for training and validation
• USA_OD_2025 reconstructs fine channels, sandbars, shoreline contours and coastal structures that are blurred in the 1/3 arc-second input
• The paper reports a mean RMSE of 0.01185, mean PSNR of 53.30 dB and mean SSIM of 0.995056 for DEM-AMSSRN
• The model reduces RMSE by up to 72.47% compared with SRGAN in the evaluated U.S. coastal tests

Data sources and preprocessing

The study selects 60 tiles from the full CUDEM collection to represent diverse East Coast, West Coast and Gulf Coast features while keeping training computationally manageable. Fifty tiles are used for training and ten for validation. Each 8112 x 8112 high-resolution training tile is divided into 96 x 96 target patches, with corresponding 32 x 32 low-resolution inputs created by bicubic downsampling at a factor of 3. A training tile yields 352,800 target patches.

Validation tiles are center-cropped to 6480 x 6480 pixels, downsampled to 2160 x 2160 pixels and divided into 60 x 60 low-resolution patches. Each validation input receives a 20-pixel mirror border, creating a 100 x 100 context window that reduces boundary artifacts during reconstruction. The elevation data use Z-score normalization to preserve local relief and avoid compressing the natural dynamic range.

Method

DEM-AMSSRN adapts an asymmetric multi-scale super-resolution network to offshore DEM reconstruction. It combines a primary feature extraction module, a nonlinear mapping module and a precision upsampling module. The nonlinear mapping stage uses region-level non-local modules, residual multi-scale blocks and residual multi-scale dilation blocks to model long-range spatial dependency, hierarchical terrain structure and enlarged receptive fields.

The residual multi-scale block is built from parallel multi-branch residual blocks. Each branch extracts local features with convolutional paths, merges the outputs with channel attention and preserves low-level information through residual connections. The residual multi-scale dilation block extends this design with dilated convolutions, enabling the network to combine local shoreline detail with broader shelf-scale context. A feature refinement fusion module combines shallow and deep features before PixelShuffle upsampling.

The training objective is a hybrid loss that combines pixel-wise L1 loss, VGG-19 perceptual loss and relativistic adversarial loss. The reported weighting coefficients are alpha = 100 for the pixel term, beta = 1 for the perceptual term and gamma = 0.1 for the adversarial term. The compared baselines are SRGAN, SRAGAN, NDSRGAN and SRADSGAN. GAN baselines use WGAN-GP, spectral normalization and adaptive learning-rate reduction when validation MSE degrades.

The main experiments use Adam with beta1 = 0.9 and beta2 = 0.999, a batch size of 64, 300 epochs and initial generator and discriminator learning rates of 1 x 10^-4. DEM-AMSSRN contains 23.63 million parameters and requires approximately 170,544 seconds for training in the reported environment, substantially less than the GAN baselines.

Validation and results

The paper evaluates RMSE, PSNR and SSIM on ten representative U.S. coastal test regions. Bicubic interpolation and four GAN-based super-resolution baselines are compared with DEM-AMSSRN at a 3x scale factor. The model reaches the lowest mean RMSE, the highest mean PSNR and the highest mean SSIM across the reported tests.

DEM-AMSSRN reconstructs clearer coastlines, sandbars, tidal channels and built structures than bicubic interpolation. On TEST5, the method preserves small tributary patterns and shoreline contours with fewer artifacts. On TEST7, it recovers narrow tidal channels, mudflats and levee alignments more continuously. Error maps for TEST8 show lower and more smoothly distributed residuals than the compared methods.

Output and access

USA_OD_2025 is a 1/9 arc-second approximately 3 m GeoTIFF offshore DEM covering major U.S. coastal zones where CUDEM source coverage is available. Dataset files are supplied through the access request form on this page. The paper, model implementation and source-data references are recorded in the project documentation for research use.